Main feature map for Claude Code

This document continues the static analysis of the extracted Claude Code runtime. Its goal is to answer a product/runtime question: what major capabilities are implemented by cli.js, and how do those capabilities connect?

The file is bundled/minified production JavaScript. The document therefore uses semantic aliases such as OuterBootstrap, CommanderRoot, HeadlessRunner, InteractiveSessionLoop, McpCoordinator, SessionRestorer, and RemoteControlBridge. Minified names are kept only as search anchors for this exact @anthropic-ai/claude-code@2.1.143 build.

Executive summary

cli.js is not a thin prompt wrapper. It is the main Claude Code agent runtime. It parses the command line, establishes process identity, loads settings and managed policy, initializes authentication and model/provider state, manages sessions, assembles tools, applies permissions, loads MCP servers and plugins, orchestrates custom agents and background agents, routes execution into interactive/headless/remote modes, and handles observability, updates, and shutdown.

Two useful lenses for the runtime are context engineering and harness engineering:

• Context engineering decides what the model can see: system prompts, CLAUDE.md, settings, output styles, tools, agents, MCP prompts/resources, memories, file inputs, and session history.
• Harness engineering decides how the model is embedded in a usable coding-agent runtime: modes, sessions, tools, permissions, hooks, streaming, retries, remote control, telemetry, and update behavior.

Source anchors

Area	Semantic alias	Minified anchor / exact string	Approximate location	Role
Bootstrap	OuterBootstrap	async function J9A	line ~19590, byte 0xddd998	Handles version fast path and lazy-loads the main bundle.
Full main	TopLevelMain	async function O4A	line ~19524, byte 0xdbe79f	Initializes environment identity, deep links, warnings, and calls w4A.
Commander	CommanderRoot	async function w4A	line ~19525, byte 0xdbf0e0	Builds root options, action body, preAction setup, and subcommands.
Headless	HeadlessRunner	async function T7A, function H89	lines ~19324-19349, bytes 0xda31bb, 0xda50d0	Runs print/SDK stream-JSON mode and drains control/message loops.
Interactive	InteractiveSessionLoop	async function pT$, async function aa4	lines ~9586-9587, bytes 0xce6582, 0xcf7dce	Runs the TUI/session loop and resume/search picker.
MCP	McpCoordinator	function fH9, function rR4	lines ~19294, ~9173	Connects runtime MCP servers and registers the mcp command tree.
Sessions	SessionRestorer	async function jHH, async function OG8	lines ~2773, ~9514	Finds recent sessions and restores transcript state.
Tools	BuiltInToolNames	Bash, Read, Edit, Write, Glob, Grep, WebFetch, WebSearch, TodoWrite, Skill	lines ~446-1091	Built-in tool-name constants and model-visible capability names.
Hooks	HookEvents	PreToolUse, PostToolUse, SessionStart, SessionEnd, SubagentStart, TaskCreated	line ~185, byte 0x10b6b4	Lifecycle hook and automation event surface.
Ops	TrafficAndDebugGates	CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC, CLAUDE_CODE_DEBUG_LOGS_DIR	lines ~62, ~133	Debug-log and telemetry/traffic policy boundaries.

Runtime system map

flowchart TB
    Cli[cli.js bundled runtime] --> Bootstrap[Bootstrap and command surface]
    Cli --> Modes[Runtime modes]
    Cli --> Context[Context and model loop]
    Cli --> Tools[Tools and permissions]
    Cli --> Sessions[Sessions and persistence]
    Cli --> Integrations[MCP / plugins / IDE / Chrome]
    Cli --> Agents[Agents and automation]
    Cli --> Ops[Diagnostics / telemetry / update]

    Bootstrap --> Commands[auth / mcp / plugin / project / agents / doctor / update / install]

    Modes --> TUI[Interactive TUI]
    Modes --> Headless[Print / SDK stream JSON]
    Modes --> Remote[Remote / teleport / Remote Control]

    Context --> Prompt[CLAUDE.md / settings / system prompt / output styles]
    Context --> Models[Auth / models / providers]
    Context --> Stream[Provider request and streaming response]

    Tools --> Builtins[Built-in coding tools]
    Tools --> Policy[Allow/deny / permission mode / hooks]
    Tools --> External[MCP tools / plugins / browser / IDE]

    Sessions --> Jsonl[local-jsonl transcripts]
    Sessions --> Resume[resume / continue / fork / rewind]
    Sessions --> Bridge[remote session ingress]

    Agents --> Custom[custom agents]
    Agents --> Task[Task and background agent commands]
    Agents --> Review[ultrareview / auto-mode]

Major feature matrix

Feature area	Entry point or trigger	Main capabilities	Primary docs
Package/Bun startup	Native Bun standalone executable, .bun graph entrypoint	Loads cli.js, pairs it with .jsc, embeds image/audio N-API modules.	Package and Bun bootstrap
CLI command shell	claude, root flags, subcommands	Version, root mode dispatch, auth, mcp, plugin, project, agents, doctor, update, install.	Commands and flags
Interactive mode	Default TTY run	Setup/login/trust screens, TUI root, resume picker, tools/agents/MCP load, and the interactive session loop.	CLI main paths
Headless/SDK mode	-p, --print, --sdk-url, non-TTY stdout, --init-only	Prompt/stdin ingestion, stream-JSON input/output, permission/control frames, JSON/text result output.	Headless streaming and resilience
Prompt/context	CLAUDE.md, .claude/settings.json, --system-prompt, --append-system-prompt, --add-dir, output styles	Runtime instruction sources, memory files, dynamic system prompt sections, slash-command/skill/agent context.	Prompt, context, and memory
Models/auth/providers	ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN, OAuth token FDs, provider env vars, --model	First-party/OAuth/API key auth, Bedrock/Vertex/Foundry/Anthropic AWS/Mantle provider selection, model aliases, thinking/budget flags.	Models, providers, and auth
Built-in tools and permissions	Tool constants plus --tools, --allowedTools, --disallowedTools, --permission-mode	File, shell, notebook, web, todo, skill, and task tools gated by filters, deny/allow rules, hooks, and permission modes.	Built-in tools and permissions
MCP/plugins/hooks	mcp, plugin, --mcp-config, --plugin-dir, hooks events	MCP config/transport/lifecycle, plugin marketplaces/session plugins, hook events and command/HTTP hook policy.	MCP, plugins, and hooks
Settings/policy/integrations	.claude/settings.json, managed settings, --settings, --ide, --chrome, statusLine	Layered settings, config roots, policy toggles, IDE/Chrome/file integration, API-key helper scripts.	Settings, policy, and integrations
Sessions and transcripts	--continue, --resume, --session-id, JSONL paths	Local transcript roots, latest-session lookup, resume/continue, fork, no-persistence, rewind.	Session resume and transcripts
Remote/teleport/control	--remote, --teleport, remote-control, --rc, remote token env vars	Remote session creation/attach, teleport hydration, Remote Control bridge, permission forwarding.	Remote control and teleport
Agents and automation	agents, --agents, task tool constants, subagent hook events, ultrareview, auto-mode	Background agents, custom agent JSON, task/subagent lifecycle, multi-agent review, permission classifier inspection.	Agents, tasks, and subagents
Diagnostics/ops/media	--debug-file, doctor, update, telemetry env vars, image/audio N-API modules	Debug logs, telemetry and traffic gates, native updater, doctor checks, media module extraction.	Diagnostics and debug logs, Telemetry and tracing, Updater and doctor, Media native modules

Takeaways

The main capabilities in cli.js can be summarized as: bootstrap, modes, context, models, tools, integrations, sessions, remote control, agents, and operations.

More concretely:

1. OuterBootstrap, TopLevelMain, and CommanderRoot form the startup and command-routing spine.
2. HeadlessRunner/HeadlessControlLoop and InteractiveSessionLoop/InteractiveResumePicker are the two main execution spines: headless/SDK and interactive TUI.
3. McpCoordinator, McpCommandRegistrar, and PluginCommandRegistrar show that MCP and plugins are first-class integration systems, not afterthoughts.
4. Tool-name constants and permission flags show a guarded action runtime around file, shell, notebook, web, todo, skill, and task capabilities.
5. SessionDiscovery, SessionRestore, JSONL transcript roots, remote tokens, bridge code, and teleport helpers show that durable sessions and remote handoff are core runtime modules.
6. Debug/telemetry/update constants and embedded image/audio N-API modules round out the operational and media-support layer.

Use this document as the map; use the linked implementation pages for source anchors and edge cases.